A customer emails in Romanian about a missed delivery. Another sends a message in Polish asking whether you can fit them in on Saturday. Until recently, many small firms would paste that into a free web translator, guess at the meaning and reply in English. Large language models such as Claude now do a noticeably better job of reading tone, context and intent, which matters more in a customer message than a word-for-word swap.
This guide covers how to use Claude to translate customer emails and messages in a UK business: what it does well, where it slips, the prompts that work, and the data protection checks to run before you paste anything in.
How many UK customers write in a language other than English?
The ONS Census 2021 language bulletin reports that 91.1% of residents aged three and over had English (or Welsh in Wales) as their main language. That leaves a sizeable minority who might prefer to write to you in Polish, Romanian, Panjabi, Urdu or another language. Most of them speak English well, so the point is not that they cannot reach you. It is that a complaint, a medical detail or a tricky question is easier to explain in your first language, and people notice when a business makes that easy.
How good is Claude at languages other than English?
Anthropic publishes a multilingual performance table that scores each language as a percentage of English performance. On that table, Claude Sonnet 4.5 scores 97.5% in French, 97.2% in Arabic, 96.7% in Hindi and 95.4% in Bengali. The smaller Claude Haiku 4.5 is lower across the board, at 92.4% in Hindi and 52.7% in Yoruba.
Two caveats are worth knowing before you read too much into those figures. First, the benchmark is a professionally translated version of a general knowledge test (MMLU), so it measures how well the model understands and reasons in each language, not how polished its translations are. Second, several of the most common UK community languages, including Polish, Romanian, Panjabi and Urdu, are not in the table at all. Anthropic's own documentation says Claude is capable in many languages beyond those benchmarked and tells you to test with the languages relevant to your use case. That is sound advice.
What does a good translation prompt look like?
Anthropic's guidance is specific here: state the input and output languages rather than relying on detection, and ask for idiomatic phrasing "as if it were a native speaker" when you are writing back. A prompt that works well for inbound messages:
"You are helping a UK plumbing firm read customer messages. Translate the following message from Polish to British English. Keep names, addresses, dates and prices exactly as written. After the translation, list anything that is ambiguous or could be read two ways, and note the customer's tone (calm, frustrated, urgent)."
For the reply, flip it:
"Translate this reply from British English to Polish. Use polite, natural phrasing a native speaker would use with a local tradesperson. Do not add or remove any commitments. Keep times in 24 hour format and prices in pounds."
The line about commitments matters. Models sometimes soften or strengthen a promise in translation, and "we will try to come Saturday" becoming "we will come Saturday" is exactly the kind of slip that turns into a complaint.
Which messages should you never translate without a human check?
A sensible rule of thumb splits your inbox into three groups:
- Send after a quick read: booking confirmations, opening hours, delivery updates, simple product questions.
- Check before sending: refunds, quotes, anything mentioning a date or price, any message where the customer sounds upset.
- Use a professional: contracts, terms and conditions, safety instructions, allergen information, anything that may end up with a solicitor or regulator.
If complaints are a regular part of your inbound mail, the approach in our guide to AI complaint handling for UK businesses pairs well with translation, because tone detection is where Claude earns its keep.
How do you handle data protection when pasting customer emails into Claude?
Under UK GDPR, a customer email almost always contains personal data, even if it is only a name and an email address. In practice that means a few habits:
- Check your plan's data settings, including whether conversations can be used for model training, and prefer a business or team plan with proper processing terms over a personal free account.
- Remove what the translation does not need. Account numbers, full card details, dates of birth and health information can usually be replaced with placeholders such as [ORDER NUMBER].
- Add AI tools to your privacy notice and record of processing activities.
- Take extra care with special category data such as health details, which carries stricter rules.
Our AI data protection checklist for UK businesses goes through each of these step by step.
Can you automate translation across your whole inbox?
The manual copy and paste approach is fine for a handful of messages a week. Past that, the time adds up, and it is easy to forget to check the ambiguity notes. An automated flow typically does four things: flags the language of each incoming email, writes an English translation into a note or label, drafts a reply in the customer's language, and holds that draft until someone clicks send.
This sits naturally alongside AI email sorting for business owners, because the same step that tags an email as a booking or a complaint can also tag it as Polish or Urdu. Keep a human approval step for the first few months at least. Removing it should be a decision you make after reviewing real results, not a default.
Should you reply in the customer's language or in English?
A bilingual reply also protects you from the rare case where a translation goes wrong. If the customer spots something odd, they can paste your English into their own tool and check. Make sure names, dates, times and prices are identical in both versions, and add a one line note such as "This reply was translated with the help of AI. The English version is below." so expectations are clear.
Which Claude model and plan suits translation work?
Anthropic's published table covers Claude Sonnet 4.5, now marked as deprecated, and Claude Haiku 4.5. It shows a clear gap between the two in lower-resource languages, with Sonnet 4.5 well ahead in Swahili and Yoruba, and a narrower gap in Spanish and French. Newer models are not in that table, so treat the pattern as a guide rather than a current ranking: if your customers mostly write in major European languages, a smaller model may be fine, and if you serve communities writing in less widely resourced languages, pay for a larger current model and keep a human check. If you are weighing costs across tools more broadly, AI customer service for small businesses in the UK covers where translation fits into a wider support setup.
What are the honest limits of AI translation for customer messages?
Customers often mix languages in one message, use transliterated text instead of the native script, or refer to a phone call you had last week. Anthropic advises submitting text in its native script for best results, so a Panjabi message written in Latin letters may translate less reliably than one written in Gurmukhi. Ask Claude to flag anything it is unsure about, and treat a list of zero ambiguities on a messy message as a warning sign, not reassurance. Used with that care, it turns a message you would have struggled with into one you can answer properly the same day.
Translate Customer Messages with Claude AI — FAQ
Can Claude AI translate customer emails accurately enough for a small business?
For routine messages such as booking changes, delivery questions and simple complaints, Claude is usually accurate enough to understand the customer and draft a clear reply. Anthropic's published multilingual benchmark puts Claude Sonnet 4.5 at 98.2% of its English performance in Spanish and 97.5% in French. That test measures understanding and reasoning, not translation quality as such. Scores drop in lower-resource languages, with Yoruba at 79.7% in the same table. Treat the output as a strong draft. Have a person check anything involving money, legal terms, medical details or a formal complaint before you send it.
Which languages should a UK business test first?
Start with the languages your customers actually write in. The ONS Census 2021 found that 8.9% of usual residents in England and Wales aged three and over, about 5.1 million people, did not have English (or Welsh in Wales) as their main language. Polish was the most common at 1.1%, followed by Romanian at 0.8%, then Panjabi and Urdu at 0.5% each. None of those four appear in Anthropic's published benchmark table, so run your own tests with real past messages before relying on Claude for them. A bilingual member of staff or a trusted customer can check a sample in an afternoon.
Is it legal under UK GDPR to paste customer emails into Claude?
It can be, but you need to treat Claude as a tool that processes personal data. That means knowing which plan you are on and its data settings, recording the use in your privacy notice and records, and removing details the translation does not need, such as account numbers, health information or full addresses. A business plan with clear data processing terms is a safer footing than a personal free account. If messages contain special category data, such as health details, take extra care and get advice. Our data protection checklist walks through the practical steps.
Should replies go back in the customer's language or in English?
Reply in the language the customer wrote in where you can, because it shows respect and reduces misunderstandings. A practical pattern is to send the translated reply with the English version underneath, so the customer can check names, dates and prices against something they can paste into their own tool. Keep your English original on file. If there is ever a dispute, you then have a record of exactly what you meant to say, separate from the machine translation.
Does Claude replace a professional translator?
No. Claude is well suited to everyday customer messages where speed matters and the cost of a small wording slip is low. Contracts, terms and conditions, safety instructions, product labels and anything that may end up in court or in front of a regulator still deserve a qualified human translator. Many businesses use Claude to triage and draft, then send the small number of high-stakes documents to a professional, which keeps translation spend focused where it matters.
